Bridged Adversarial Training

TL;DR

This paper introduces Bridged Adversarial Training, a novel method that improves robustness of neural networks by addressing the gap between clean and adversarial examples, supported by theoretical and empirical evidence.

Contribution

It proposes a new adversarial training approach that mitigates the negative effects of smoothness regularizers and enhances robustness against large perturbations.

Findings

Bridged adversarial training improves robustness stability.

The method outperforms existing techniques on large perturbations.

Theoretical analysis supports empirical results.

Abstract

Adversarial robustness is considered as a required property of deep neural networks. In this study, we discover that adversarially trained models might have significantly different characteristics in terms of margin and smoothness, even they show similar robustness. Inspired by the observation, we investigate the effect of different regularizers and discover the negative effect of the smoothness regularizer on maximizing the margin. Based on the analyses, we propose a new method called bridged adversarial training that mitigates the negative effect by bridging the gap between clean and adversarial examples. We provide theoretical and empirical evidence that the proposed method provides stable and better robustness, especially for large perturbations.